One may ask why we are writing so much about the AI–human intersection in complex contexts.
The simple answer is: we are learning too.
We are learning what will actually be useful in real environments, especially when it comes to what we might call context intelligence — the ability to understand not just information, but the environment, the system, and the situation that information must live inside.
And in this journey, we came across something interesting in AI called overfitting.
Overfitting is when an AI (or a model) learns something too perfectly from its practice examples, so it starts to memorize instead of understand. Yes, I said understand.
Because understanding is the key difference.
A simple way to see it:
- Good learning: “Cats have whiskers, pointy ears, and fur.”
- Overfitting: “A cat is ONLY this exact orange cat sitting on my couch.”
So when the model sees a black cat or a fluffy cat, it gets confused and says, “That’s not a cat!”
Now let me speak more human for a moment.
If you learn the idea of spelling rules, you can spell new words you’ve never seen before. But if you only memorize the exact practice sheet, you might do great on that sheet — perfect even — but then fail the real test when the words are different.
That is the risk of overfitting.
And interestingly, this is not just an AI problem. It shows up in human systems too. Nowhere more clearly than in medicine.
When the Standard of Healing Travels
Medicine is supposed to be universal. A bacterium is a bacterium. A fracture is a fracture. The human body, more or less, is the human body. And yet anyone who has worked across health systems knows that the standards of healing developed in one context routinely fail sometimes catastrophically when they are picked up and dropped into another.
This is overfitting, but in white coats.
Consider the management of childhood diarrhoea. In a well-resourced hospital in Toronto or London, the standard of care is clear: assess the child, run electrolyte panels, start IV fluids if the child is severely dehydrated, monitor with continuous observation, and adjust. The protocol assumes a lab down the hall, a paediatric nurse per few beds, reliable electricity, sterile IV sets in supply, and a parent who can stay at the bedside without losing a week’s income.
Now move that same protocol to a rural district hospital in northern Uganda or eastern DRC. The child arrives after a six-hour journey on the back of a motorbike. The lab is closed for the weekend. The IV stock ran out on Tuesday. The “continuous monitoring” is one nurse covering forty beds. The mother is holding two other children and has no one at home to feed them.
The Toronto protocol, applied faithfully, does not heal this child. In some cases it actively delays the thing that would — oral rehydration solution, mixed correctly, given patiently by the mother, with zinc supplementation and a clear sign of when to escalate. The WHO knew this decades ago. ORS saves more children globally than almost any other intervention in history precisely because it was designed for the context, not transplanted into it.
The clinician who insists on the Toronto standard in the Ugandan ward is overfitting. They learned medicine on one dataset and cannot generalise to another.
The Same Disease, Two Different Diseases
Take tuberculosis. In a high-income setting, TB is largely a story of latent infection, individual cases, drug-sensitive strains, and patients who can be expected to complete six months of supervised therapy because their lives are stable enough to do so. The standard of healing is: diagnose with GeneXpert, prescribe the regimen, follow up, cure.
In a crowded urban informal settlement in Nairobi or Mumbai, TB is a different disease entirely, even though the bacterium is identical. It is a story of household transmission in single-room dwellings, of patients who disappear from follow-up because they had to migrate for work, of multidrug resistance born from interrupted courses, of co-infection with HIV, of stigma so heavy that diagnosis itself is avoided. The same six-month regimen, prescribed identically, produces wildly different outcomes — not because the drug is different, but because the context the drug must live inside is different.
A doctor trained only on the first version of TB, parachuted into the second, will write the right prescription and watch the wrong outcome unfold. The medicine was correct. The context intelligence was missing.
Maternal Bleeding, Two Standards
Postpartum haemorrhage is the leading cause of maternal death globally. In a tertiary hospital in Calgary, the standard of healing involves rapid blood transfusion, uterotonic drugs on hand, an obstetric team in the room within minutes, an operating theatre two doors away, and an interventional radiology suite if it comes to that.
In a health post three hours from the nearest road in highland Ethiopia, the standard of healing for the same condition has to be reimagined from the ground up. It is misoprostol distributed in advance to traditional birth attendants. It is the non-pneumatic anti-shock garment to buy time. It is uterine balloon tamponade improvised from a condom and a catheter — yes, really, and it works. It is a referral system designed around motorbike ambulances and community phone trees.
A clinician who knows only the Calgary standard will, in the Ethiopian highland, watch a woman die while waiting for resources that are not coming. A clinician with context intelligence will reach for the misoprostol and the condom catheter and save her.
Same condition. Same physiology. Two completely different standards of healing — and both are correct in their context.
What Medicine Teaches Us About Overfitting
The pattern repeats across every domain of health. Mental health protocols built around weekly one-hour therapy sessions do not survive contact with communities where healing has always been collective and ceremonial. Nutritional guidance built around supermarket access collapses where the food system is seasonal and subsistence. Surgical checklists designed for theatres with running water need rethinking where the water comes in a jerrycan.
None of this means the original protocols are wrong. They are correct for the context in which they were learned. The error is in assuming the protocol is the medicine. The protocol is one expression of the medicine, shaped by one environment. The medicine itself — the principle, the physiology, the goal of healing — has to be re-expressed in each new context.
This is the medical version of what AI researchers call generalisation. A well-generalised model holds onto the underlying principle and lets the specifics flex. An overfitted model holds onto the specifics and breaks when they change.
When Humans Overfit
It is tempting to think of overfitting as something that happens to machines. It does not. It happens to us — daily, professionally, institutionally — and often by people who are very good at what they do. A few familiar instances:
The surgeon who only knows one technique. A surgeon trained in a single hospital, on a single operating table, with one anaesthetist and one set of instruments, can become extraordinary in that room and lost outside it. Move them to a facility where the suction does not work, the lights flicker, and the assistant has never seen the procedure, and a brilliant operator becomes hesitant. The skill was real; it was just bonded too tightly to one environment.
The expat manager who runs the same playbook everywhere. A development professional builds a successful programme in Bangladesh — community mobilisation, women’s savings groups, a particular monitoring rhythm. They are then posted to South Sudan and run the same playbook. The savings groups do not form. The mobilisation meetings are seen as politically suspect. The monitoring forms ask questions that do not map to how people live. The playbook was not wrong in Dhaka. It was overfitted to Dhaka.
The clinician who cannot read the patient in front of them. A doctor who has only ever practised in a teaching hospital learns to trust the lab over the patient. Numbers, scans, panels. Drop that same clinician in a setting where the labs are unreliable or unavailable, and they freeze — because the clinical examination, the patient’s story, the family’s account, the look in the eyes of the mother, were never trusted as primary data. They overfitted on instruments and lost the older, deeper skill of seeing.
The teacher who teaches to the test they were taught by. A teacher who succeeded in a system of rote memorisation often teaches the next generation to memorise — even when the students they now face will need to navigate ambiguity, judgement, and contexts that have no answer key. The teacher is not lazy. They are running the model that produced them.
The policymaker who imports a foreign template. A ministry official returns from a study tour and proposes a national health insurance scheme modelled on Germany or Rwanda. The architecture is elegant. It assumes a tax base, a formal employment sector, a registry of citizens, and a culture of trust in central institutions. None of which hold in the country it is being imported into. The policy is not bad policy. It is overfitted policy.
The aid worker who built one famine response and now builds every famine response the same way. The 2011 Somalia response, the 2017 South Sudan response, the 2022 Horn of Africa response — different shocks, different governance, different displacement patterns, different markets. But the muscle memory of the first response keeps reaching for the same playbook. Sometimes it works. Often the gaps it misses are exactly the things that made this crisis different from the last one.
The leader who keeps running the strategy that built the company. A founder who scaled a business through aggressive hiring and a particular culture of hustle keeps running that strategy long after the company, the market, and the workforce have changed. The strategy that built the first hundred employees actively damages the next thousand. The leader is not stubborn — they are pattern-matching on the only environment in which they have ever succeeded.
In every one of these cases, the person is competent. Often, they are the best in their field. The failure is not of skill. It is of transfer. They learned deeply in one context and assumed the learning was the context-free truth. It was not. It never is.
This is the quiet, unglamorous discipline that context intelligence demands: knowing the difference between what you learned and where you learned it, and being willing to question the second every time you move.
Why This Matters Now
When we write about the AI–human intersection in complex contexts, this is the territory we are mapping. Because AI systems are now being deployed into health, finance, agriculture, education, and governance across exactly the kinds of contexts where overfitting causes the most damage — places where the training data came from somewhere else and the local reality does not match.
An AI triage tool trained on Boston emergency department data will overfit when it lands in Kampala. A credit-scoring model trained on US consumer data will overfit when it lands in rural Bangladesh. A crop disease classifier trained on California fields will overfit when it lands in a smallholder plot in Malawi.
The fix is not to abandon the tools. The fix is context intelligence — the deliberate, structured practice of asking what changes when the context changes, and being willing to rebuild the standard of healing accordingly.
This is why context matters. And why learning is not just about repetition — it is about transfer.
